agentic-delegation

# Agentic Delegation The orchestrator's context window is the scarcest resource in a session. Every file dump, log tail, and grep result read directly is context that can no longer hold the plan, the spec, or the diff under review. Delegation exists to keep that window for judgment. Opus 5.5 is the daily driver: it is the orchestrator, and it makes the code edits. Reading work goes to cheaper models. ## Route each step by the kind of work | Work | Who does it | Model | Effort | | --- | --- | --- | --- | | Search, recon, mapping call sites, "where does X live" | subagent (`Explore`) | `haiku`; `sonnet` when interpreting what it finds takes judgment | none on Haiku | | Reading CI or gate logs, running the gate suite | subagent | `haiku` or `sonnet` | none on Haiku | | Bulk prose: docs pages, guides, templates, against an exact spec | subagent | `sonnet` | session level | | Code edits: scripts, tests, generators, gates, workflow code, Rust | orchestrator, or `opus` subagents for independent parallel edits | Opus 5.5 | session level | | Architecture, schema shapes, security-sensitive and load-bearing edits | orchestrator only | Opus 5.5 | session level | | Verification of delegate output, and the commit | orchestrator only | Opus 5.5 | session level | | One task that `high` has failed twice in the same way | orchestrator, after switching model | Fable 5.1 | then switch back | Search and log-reading are summarization jobs: a cheaper model reads a lot and returns a little, and...